DETAILED ACTION
Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
All objections/rejections not mentioned in this Office Action have been withdrawn by the Examiner.
Examiner’s Note
Applicant’s arguments regarding the structural relationship between the “relevant digital documents,” the “ground truth digital documents,” and the “question-answer pairs” have been fully considered. Further, it is understood that the applicant is relying on the “question-answer pairs” to distinguish the cited references regarding claim 9. However, it is respectfully noted that applicant’s usage of “question-answer pairs” in the specification creates a significant ambiguity between standard machine learning terminology and applicant’s implied usage of these terms, rendering the structural boundaries of these claims unclear.
As understood to a person of ordinary skill in the art of supervised machine learning and retrieval augmented generation (RAG), a “document” serves as the passive source context, while a “question-answer pair” serves as the active training label (the input query and the target output). When performing supervised training of a generative model and calculating loss, the target answer label is computationally isolated from the source document during the model’s input phase, even if they are stored or bundled together for training. This computational isolation is necessary to avoid catastrophic data leakage. The specification, however, blurs this well-understood paradigm by grammatically conflating the source context with the training labels.
This confusion is compounded by contradictory descriptions of the data architecture within the specification. On one hand, the specification implies a nested physical structure by referencing “ground truth documents (including a question-answer pair).” (Instant Application, ¶ [0019]). On the other hand, the specification teaches using a large language model (LLM) to synthetically generate question-answer pairs from raw help video transcripts and help documents (Instant Application, ¶ [0053]). Further, the specification indicates that, during the supervised training phase, the system generates question-answer pairs “for the ground truth documents.” (Instant Application, ¶ [0097]). A label algorithmically generated for or from a document is understood by a person having ordinary skill in the art to be an external, derived data structure, not a native physical string housed within the original document.
Rather than acting as their own lexicographer to clearly delineate any intended separation, the specification doesn’t appear to recognize the distinction between the two distinct, well-established meanings of “question-answer pairs” within the art. First, the specification appears to utilize the phrase in the context of supervised machine learning, where a question-answer pair represents an input prompt and a withheld target label used strictly for loss calculation (e.g., separating the answer from the model’s input to prevent data leakage). Second, the specification appears to utilize the phrase to describe a literal text formatting convention within a source document, such as a FAQ page where a question and an answer are printed sequentially in the text payload. However, unfortunately, the specification never expressly recognizes the two implied meanings, such that related portions of the disclosure can be understood in light of that relationship. By using a single phrase interchangeably where said phrase could be understood as both a discrete, isolated supervised training label and a native text formatting convention within a single data pipeline (as visually conflated in FIG. 7), the specification of the instant application creates an internal conflict that renders the structural boundaries of the claims directed to that limitation unclear.
Due to the integrated nature of the problem in the specification, a reliable road map to avoiding and/or overcoming 112(a) and 112(b) rejections in this instance is not readily clear. However, given the two equally plausible and intermingled definitions for the phrase, clear indications of the intended meaning of “question-answer pairs,” as presented or utilized in amendments to the claims, and a clear indication of the specification support relied on in the amendment, will help avoid future rejections on lack of clarity and lack of written description.
Status of the Claims
Prior to entry of the amendment(s) and/or consideration of the argument(s), the status of the claims is as follows.
Claim(s) 1-20 is/are pending.
Claim(s) 1, 2, 8, and 15-16 is/are rejected under 35 U.S.C. 102(a)(2) as being anticipated by Subbiah (U.S. Pat. App. Pub. No. 2025/0156455, hereinafter Subbiah).
Claims 3 is/are rejected under 35 U.S.C. 103 as being unpatentable over Subbiah as applied to claim 1 above, and further in view of Kar (U.S. Pat. App. Pub. No. 2025/0190802, hereinafter Kar).
Claims 5-7, and 17-18 is/are rejected under 35 U.S.C. 103 as being unpatentable over Subbiah as applied to claim 1 above, and further in view of Chen (U.S. Pat. App. Pub. No. 2025/0077792, hereinafter Chen).
Claims 9, 11, and 14 is/are rejected under 35 U.S.C. 103 as being unpatentable over Subbiah in view of Kar and Chen.
Claims 10 is/are rejected under 35 U.S.C. 103 as being unpatentable over Subbiah, Kar, and Chen as applied to claim 9 above, and further in view of Joachims.
Claims 12 is/are rejected under 35 U.S.C. 103 as being unpatentable over Subbiah, Kar, and Chen as applied to claim 9 above, and further in view of Zhan (CN112836068A, hereinafter Zhan).
Claims 13 is/are rejected under 35 U.S.C. 103 as being unpatentable over Subbiah, Kar, and Chen as applied to claim 9 above, and further in view of Non-Patent Literature to Xiong (Xiong, L., Xiong, C., Li, Y., Tang, K.F., Liu, J., Bennett, P., Ahmed, J. and Overwijk, A., 2020. Approximate nearest neighbor negative contrastive learning for dense text retrieval. arXiv preprint arXiv:2007.00808, hereinafter Xiong).
Claims 19 is/are rejected under 35 U.S.C. 103 as being unpatentable over Subbiah as applied to claim 15 above, and further in view of Chen and Mahajan (U.S. Pat. App. Pub. No. 2021/0026820, hereinafter Mahajan).
Response to Amendments
Applicant’s amendment filed on 22 January 2026 has been entered.
In view of the amendment to the claim(s), the amendment of claim(s) 1, 8-9, 11, and 15 have been acknowledged and entered.
In view of the amendment to claim(s) 1, 8-9, 11, and 15, the rejection of claims 1-20 under 35 U.S.C. §102 and 103 is withdrawn.
In light of the amended claims, new grounds for rejection under 35 U.S.C. §112 and 35 U.S.C. §103 are provided in the action below.
Response to Arguments
Applicant’s arguments regarding the prior art rejections under 35 U.S.C. §102/103, see pages 13-18 of the Response to Non-Final Office Action dated 17 November 2025, which was received on 22 January 2026 (hereinafter Response and Office Action, respectively), have been fully considered.
With respect to the rejection(s) of claim(s) 1 and 15 under 35 U.S.C. §102 as being anticipated by Subbiah, applicant asserts that Subbiah fails to teach or suggest all limitations of amended claims 1 and 15. Applicant’s arguments are persuasive. Therefore, the rejection of claims 1 and 15 are withdrawn.
With respect to the rejection(s) of claim(s) 9 under 35 U.S.C. §103 in light of Subbiah in view of Subbiah in view of Kar and Chen, applicant asserts that the cited references fail to teach or suggest all limitations of amended claim 9. Applicant’s arguments are persuasive. Therefore, the rejection of claim 9 is withdrawn.
Applicant further argues that the rejection(s) of dependent claims 2-8, 10-14, and 16-20 should be withdrawn for at least the same reasons as independent claims 1, 9, and 15. Applicant’s arguments in light of the amended claims are persuasive. As such, the rejections of claims 2-8, 10-14, and 16-20 under 35 U.S.C. §102 and 35 U.S.C. §103 are withdrawn.
However, upon further consideration, new ground(s) of rejection under 35 U.S.C. §103 are made in light of combinations of Subbiah, Kar, Chen, Joachims, Zhan, Xiong, and newly cited references Non-Patent Literature to Katsogiannis-Meimarakis (Katsogiannis-Meimarakis, G. and Koutrika, G., 2023. A survey on deep learning approaches for text-to-SQL. The VLDB Journal, 32(4), pp.905-936, hereinafter Katsogiannis) and Non-Patent Literature to Siriwardhana (Siriwardhana, S., Weerasekera, R., Wen, E., Kaluarachchi, T., Rana, R. and Nanayakkara, S., 2023. Improving the domain adaptation of retrieval augmented generation (RAG) models for open domain question answering. Transactions of the Association for Computational Linguistics, 11, pp.1-17., hereinafter Siriwardhana).
The Applicant has not provided any further statement and therefore, the Examiner directs the Applicant to the below rationale.
Claim Rejections - 35 USC § 112
The following is a quotation of the first paragraph of 35 U.S.C. 112(a):
(a) IN GENERAL.—The specification shall contain a written description of the invention, and of the manner and process of making and using it, in such full, clear, concise, and exact terms as to enable any person skilled in the art to which it pertains, or with which it is most nearly connected, to make and use the same, and shall set forth the best mode contemplated by the inventor or joint inventor of carrying out the invention.
The following is a quotation of the first paragraph of pre-AIA 35 U.S.C. 112:
The specification shall contain a written description of the invention, and of the manner and process of making and using it, in such full, clear, concise, and exact terms as to enable any person skilled in the art to which it pertains, or with which it is most nearly connected, to make and use the same, and shall set forth the best mode contemplated by the inventor of carrying out his invention.
Claims 9-14 are rejected under 35 U.S.C. 112(a) or 35 U.S.C. 112 (pre-AIA ), first paragraph, as failing to comply with the written description requirement. The claim(s) contains subject matter which was not described in the specification in such a way as to reasonably convey to one skilled in the relevant art that the inventor or a joint inventor, or for applications subject to pre-AIA 35 U.S.C. 112, the inventor(s), at the time the application was filed, had possession of the claimed invention.
Regarding claim 9, claim 9 recites an architecture which lacks specification support. Specifically, the unlinked data pipeline described in amended claim 9 is not supported by the specification as filed. Claim 9 recites modifying parameters of a context retrieval model to learn relevant digital documents “comprising relevant question answer pairs” and subsequently recites generating a response from a “ground truth digital document… wherein the ground truth digital document comprises one or more question answer pairs.” Because the “one or more question answer pairs” lacks antecedent basis linking it to the “relevant question answer pairs,” the claim encompasses an embodiment wherein the response generator model trains on a completely separate set of question-answer pairs that are not unlinked to the relevant answer pairs initially retrieved by the context retrieval model.
The specification does not describe or otherwise contain support for an embodiment where the context retrieval model and the response generator model operate on disconnected, structurally separate lineages of question-answer pairs. The description in the specification is limited to a cohesive pipeline where the generator model processes relevant data derived directly from the retrieval phase. The omission of the structural linkage between these datasets in the claim language, results in a set of limitations which are not supported in the specification of the instant application at the time of filing. Therefore, claim 9 contains limitations which fail for lack of written description and is therefore rejected under 112(a).
Regarding claims 10-14, claims 10-14 depend from claim 9 and incorporate all limitations therefrom. Therefore, claims 10-14 are rejected under 112(a) for at least the same reasons as described above with relation to claim 9.
The following is a quotation of 35 U.S.C. 112(b):
(b) CONCLUSION.—The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the inventor or a joint inventor regards as the invention.
The following is a quotation of 35 U.S.C. 112 (pre-AIA ), second paragraph:
The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the applicant regards as his invention.
Claims 9-14 are rejected under 35 U.S.C. 112(b) or 35 U.S.C. 112 (pre-AIA ), second paragraph, as being indefinite for failing to particularly point out and distinctly claim the subject matter which the inventor or a joint inventor (or for applications subject to pre-AIA 35 U.S.C. 112, the applicant), regards as the invention.
Regarding claim 9, the limitation “question-answer pairs” lacks clarity. Claim 9 recites “question answer pairs” at lines 5, 8, and 13, respectively. Though each of these claim parts rely on the same root phrase “question-answer pairs”, the claim parts do not have antecedent basis and do not indicate a relationship to one another. As such, it is unclear how the “question-answer pairs” of the “stored digital documents” relate to the “relevant question answer pairs” of the “relevant digital documents”, or how either of the previous two relate to the “one or more question-answer pairs” of the “ground truth digital document”. This confusion is further compounded by lack of necessary connection between the training dataset (as generated from the “stored digital documents) and either of the “relevant digital documents” or the “ground truth digital document.” As well, given the unclear usage of “question-answer pairs” discussed in the examiner note above, the context in which each of the “question-answer pairs” is being applied is also in question. Therefore, the limitation “question-answer pairs” lacks clarity and claim 9 is rejected.
If these are understood to be the same claim part, applicant should revise to clarify the antecedent basis between the above parts or otherwise indicate their relationship (e.g., through group and sub-group indications). If these are to be considered different parts, applicant is advised to amend the limitations such that the claim parts are clearly separate or otherwise do not rely on the same root part name.
Regarding claims 10-14, claims 10-14 depend from claim 9 and incorporate all limitations therefrom. Therefore, claims 10-14 are rejected under 112(b) for at least the same reasons as described above with relation to claim 9.
Appropriate correction is required.
Claim Rejections - 35 USC § 103
In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis (i.e., changing from AIA to pre-AIA ) for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
Claim(s) 1, 2, 8, and 15-16 is/are rejected under 35 U.S.C. 103 as being unpatentable over Subbiah in view of Katsogiannis.
Regarding claim 1, Subbiah discloses A computer-implemented method comprising (Systems and methods described with reference to “using generative AI to process user queries in the construction domain”; Subbiah, ¶ [0105]): receiving a contextual query comprising terminology indicating a software context of a computer application (“At step 302, a user query is received/obtained (e.g., entered into a text dialog box and/or selected from a list of predefined queries) within a construction software system,” and the terminology can indicate a software context of a computer application (e.g., requests for information of RFIs received with regards to a construction application.); Subbiah, ¶ [0055], [0106]); mapping, using a context mapping model, the terminology to a software-specific domain… (“At step 304, a user query processing component may perform pre-processing of the user (NL) query including correcting typos, verifying the quality and safety of user input, and rephrasing and complementing user queries (i.e., to validate the query)” where “rephrasing and complementing user queries...to validate the query” is understood as resolving terminology in the query {contextual query}, and where the “rephrasing and complementing” includes “modify[ing] the query 402” such as by “generat[ing] general SQL queries that access relational datasets {...to a software-specific domain}”; Subbiah, ¶ [0107], [0119], [0130]); extracting, using a context retrieval model trained to determine relevant contextual data for software contexts of a plurality of computer applications, a query embedding from the contextual query and a plurality of data segment embeddings from data segments of stored digital documents (“At step 306, an embedding model may then be utilized to embed the pre-processed text from the query into vectors for the semantic search” and “the semantic search is performed within each of the multiple different sections based on the search vectors and the stored vectors.”; Subbiah, ¶ [0108], [0111]); determining, using the context retrieval model, relevant digital documents from among the stored digital documents by comparing the query embedding and the plurality of data segment embeddings (“The semantic search identifies the semantically relevant sections of the multiple different sections (i.e., the semantic search filters the data in the data source to improve the search quality and accuracy)” by comparison between the “search vector and the stored vectors”; Subbiah, ¶ [0111]); and generating, from the relevant digital documents, a contextual response for the contextual query (“an LLM is utilized to generate a response based on the contextual data prompt,” where the contextual data prompt includes “the semantically relevant sections of the multiple different sections (i.e., the semantic search filters the data in the data source to improve the search quality and accuracy)” and the embeddings of “the pre-processed text from the query”; Subbiah, ¶ [0108], [0111], [0113]) utilizing a response generator model tuned to generate context-specific responses for the software contexts of the plurality of computer applications (Generative model as directed to “process[ing] user queries in the construction domain...within a construction software system” are fine-tuned and trained “to become more attuned to the construction industry’s evolving landscape” where the generative mode, which may also be a LLM, “the LLM is trained based on construction data and the response identifies the semantically relevant section(s)”; Subbiah, ¶ [0029], [0105]-[0106], [0108], [0113], [0120]). However, Subbiah fails to expressly recite mapping, using a context mapping model, the terminology to a software-specific domain by resolving ambiguous terminology in the contextual query to context-specific terminology associated with the software context of the computer application.
Katsogiannis teaches systems and methods for “deep learning text-to-SQL.” (Katsogiannis, p. 907, col. 1, lines 36-40). Regarding claim 1, Katsogiannis teaches mapping, using a context mapping model, the terminology to a software-specific domain (Discloses “natural language representation” for “text-to-SQL systems” and relational database schema where relational database schema is a software-specific domain, and the mapping of natural language to that relational database schema is the mapping of terminology to that software-specific domain, and where the exemplary models such as RAT-SQL and various PLMs (BERT, RoBERTa, etc.) are context mapping models.; Katsogiannis, ¶ p. 914, col. 1, lines 45- col. 2, lines 8; p. 914-915, section 4.2.2) by resolving ambiguous terminology in the contextual query to context-specific terminology associated with the software context of the computer application (Sections 2. and 2. discuss the mechanics of disambiguation from the perspective of the natural language and from the perspective of the SQL query generated therefrom. Resolving the inherent ambiguities in natural language input includes “context-specific ambiguity {resolving ambiguous terminology...}” of natural-language query {the context-specific query...} due to “a term having different meanings depending on the query context {to context-specific terminology...}, the data domain {...associated with the software context of the computer application}, and the user goals” where, as explained with relation to an example, the word “top” includes both user intent context and software domain context in determining that “the query “Return the top movie” on a movie database, “top” may mean based on the number of ratings collected” and based on “SQL...[having] a strict syntax” the system resolves “the query ‘Return the movie with the best rating’... to a nested SQL query,” where resolving ambiguous terminology to context-specific terminology for text-to-SQL includes resolve the “vocabulary gap” (e.g., that the word “actress” resolves to “Actor.name”, where actress is ambiguous in the context of a database which does not rely on gender specific descriptors) and “schema ambiguity” (e.g., logically resolving whether “model” refers to “car.model” or “engine.model”).; Katsogiannis, ¶ p. 907-908, sections 2.1 and 2.2).
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the two-stage retrieval augmented generation of Subbiah to incorporate the teachings of Katsogiannis to include mapping, using a context mapping model, the terminology to a software-specific domain by resolving ambiguous terminology in the contextual query to context-specific terminology associated with the software context of the computer application. Subbiah discloses modification of a semantic query prior to generating the embedding (steps 304 and 306) to improve semantic search. Katsogiannis discloses that, when processing natural language for database queries, resolving ambiguity involves modifying the query string itself (via input serialization) to append schema-linking tokens or database-specific values, aligning the terminology with the database schema prior to generating the embeddings. It would have been obvious to modify the disclosure of Subbiah with the schema aligned input serialization taught by Katsogiannis, as part of a pre-processing step of step 304, to tag and align ambiguous terminology with strict relational database schema of the BDP structured data, thereby enriching the semantic vectors generated in step 306 and ensuring the downstream SQL query engine formulates highly accurate, error-free database commands, as recognized by Katsogiannis. (Katsogiannis, p. 907, col. 1, lines 1-18).
Regarding claim 2, the rejection of claim 1 is incorporated. Subbiah and Katsogiannis disclose all of the elements of the current invention as stated above. Subbiah further discloses wherein determining the relevant digital documents comprises: determining a context score for a stored digital document from among the stored digital documents based on comparing the query embedding and the plurality of data segment embeddings (“an optional information retrieval strategy is utilized to optimize information retrieval (e.g., in order to conduct a reliable semantic search)” which can include “a similarity score (e.g., utilized for measuring the text relevancy for the information retrieval)”; Subbiah, ¶ [0110]); and selecting the stored digital document as a relevant digital document based on the context score (In the optional embodiment, the similarity score is “utilized for measuring the text relevancy for the information retrieval” and the most relevant sections/text is selected; Subbiah, ¶ [0110]).
Regarding claim 8, the rejection of claim 1 is incorporated. Subbiah and Katsogiannis disclose all of the elements of the current invention as stated above. Subbiah further discloses wherein resolving ambiguous terminology in the contextual query (As explained above, the mapping of terminology to a software specific domain occurs at step 304 of Subbiah, where the “rephrasing and complementing user queries...to validate the query” is modified by the disambiguation techniques described in Katsogiannis in sections 2. and 2., to “resolving ambiguous terminology in the contextual query” as described above with relation to claim 1; Subbiah, ¶ [0107], [0119], [0130]) prior to generating the query embedding (“At step 306, an embedding model may then be utilized to embed the pre-processed text from the query into vectors for the semantic search” and “the semantic search is performed within each of the multiple different sections based on the search vectors and the stored vectors.” As such, the embedding occurs after the preprocessing {...prior to generating the query embedding}; Subbiah, ¶ [0108], [0111]). However, Subbiah fails to expressly recite mapping, using a context mapping model, the terminology to a software-specific domain by resolving ambiguous terminology in the contextual query to context-specific terminology associated with the software context of the computer application.
The relevance of Katsogiannis is described above with relation to claim 1. Regarding claim 8, Katsogiannis teaches comprises modifying the contextual query to align the terminology with a software-specific domain representation (With relation to the explanation above, the system of Katsogiannis includes mapping the “strict syntax” of SQL to the natural language input including resolve the “vocabulary gap” (e.g., that the word “actress” resolves to “Actor.name”, where actress is ambiguous in the context of a database which does not rely on gender specific descriptors). Resolving the vocabulary gap, as described in the context of this example, is a modification of at least one word (e.g., actress) of the contextual query {comprises modifying the contextual query} to “actor”, to align the terminology with a software-specific domain representation (e.g., “actor.name”).; Katsogiannis, ¶ p. 907-908, sections 2.1 and 2.2).
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the two-stage retrieval augmented generation of Subbiah to incorporate the teachings of Katsogiannis to include mapping, using a context mapping model, the terminology to a software-specific domain by resolving ambiguous terminology in the contextual query to context-specific terminology associated with the software context of the computer application. Subbiah discloses modification of a semantic query prior to generating the embedding (steps 304 and 306) to improve semantic search. Katsogiannis discloses that, when processing natural language for database queries, resolving ambiguity involves modifying the query string itself (via input serialization) to append schema-linking tokens or database-specific values, aligning the terminology with the database schema prior to generating the embeddings. It would have been obvious to modify the disclosure of Subbiah with the schema aligned input serialization taught by Katsogiannis, as part of a pre-processing step of step 304, to tag and align ambiguous terminology with strict relational database schema of the BDP structured data, thereby enriching the semantic vectors generated in step 306 and ensuring the downstream SQL query engine formulates highly accurate, error-free database commands, as recognized by Katsogiannis. (Katsogiannis, p. 907, col. 1, lines 1-18).
Regarding claim 15, Subbiah discloses A system comprising: one or more memory devices; and one or more processors coupled to the one or more memory devices (Systems and methods described with reference to “using generative AI to process user queries in the construction domain” which may be “tangibly embodied in a non-transitory computer-readable medium” as part of a “computer 902” including a “processor 904” and a “memory 906” coupled to said processor 904, where “the operating system 908 and the computer program 910 are comprised of computer program 910 instructions which, when accessed, read and executed by the computer 902, cause the computer 902 to perform the steps necessary to implement and/or use the present invention”; Subbiah, ¶ [0105], [0140], [0147]), the one or more processors configured to cause the system to: receive a contextual query comprising terminology indicating a software context of a computer application (“At step 302, a user query is received/obtained (e.g., entered into a text dialog box and/or selected from a list of predefined queries) within a construction software system,” and the terminology can indicate a software context of a computer application (e.g., requests for information of RFIs received with regards to a construction application.); Subbiah, ¶ [0055], [0106]); map, using a context mapping model, the terminology to a software-specific domain… (“At step 304, a user query processing component may perform pre-processing of the user (NL) query including correcting typos, verifying the quality and safety of user input, and rephrasing and complementing user queries (i.e., to validate the query)” where “rephrasing and complementing user queries...to validate the query” is understood as resolving terminology in the query {contextual query}, and where the “rephrasing and complementing” includes “modify[ing] the query 402” such as by “generat[ing] general SQL queries that access relational datasets {...to a software-specific domain}”; Subbiah, ¶ [0107], [0119], [0130]);; extract, using a context retrieval model trained to determine relevant contextual data for software contexts of a plurality of computer applications: a query embedding from the contextual query (“At step 306, an embedding model may then be utilized to embed the pre-processed text from the query into vectors for the semantic search” and “the semantic search is performed within each of the multiple different sections based on the search vectors and the stored vectors.”; Subbiah, ¶ [0108], [0111]); and a plurality of data segment embeddings by determining semantically consistent segments of stored digital documents and extracting embeddings from the semantically consistent segments (“The data source has multiple different sections and each data source is a vector database with stored vectors representing the multiple different sections” where “vector store may be utilized as a dedicated repository for storing embedding vectors that are associated with specification chunks.”; Subbiah, ¶ [0109], [0122]); determine, utilizing the context retrieval model, relevant digital documents from among the stored digital documents by comparing the query embedding and the plurality of data segment embeddings (“The semantic search identifies the semantically relevant sections of the multiple different sections (i.e., the semantic search filters the data in the data source to improve the search quality and accuracy)” by comparison between the “search vector and the stored vectors”; Subbiah, ¶ [0111]); and generate, from the relevant digital documents, a contextual response for the contextual query (“an LLM is utilized to generate a response based on the contextual data prompt,” where the contextual data prompt includes “the semantically relevant sections of the multiple different sections (i.e., the semantic search filters the data in the data source to improve the search quality and accuracy)” and the embeddings of “the pre-processed text from the query”; Subbiah, ¶ [0108], [0111], [0113]) utilizing a response generator model tuned to generate context-specific responses for the software contexts of the plurality of computer applications (Generative model as directed to “process[ing] user queries in the construction domain...within a construction software system” are fine-tuned and trained “to become more attuned to the construction industry’s evolving landscape” where the generative model, which may also be a LLM, “the LLM is trained based on construction data and the response identifies the semantically relevant section(s)”; Subbiah, ¶ [0029], [0105]-[0106], [0108], [0113], [0120]). However, Subbiah fails to expressly recite map, using a context mapping model, the terminology to a software-specific domain by resolving ambiguous terminology in the contextual query to context-specific terminology associated with the software context of the computer application.
The relevance of Katsogiannis is described above with relation to claim 1. Regarding claim 15, Katsogiannis teaches mapping, using a context mapping model, the terminology to a software-specific domain (Discloses “natural language representation” for “text-to-SQL systems” and relational database schema where relational database schema is a software-specific domain, and the mapping of natural language to that relational database schema is the mapping of terminology to that software-specific domain, and where the exemplary models such as RAT-SQL and various PLMs (BERT, RoBERTa, etc.) are context mapping models.; Katsogiannis, ¶ p. 914, col. 1, lines 45- col. 2, lines 8; p. 914-915, section 4.2.2) by resolving ambiguous terminology in the contextual query to context-specific terminology associated with the software context of the computer application (Sections 2. and 2. discuss the mechanics of disambiguation from the perspective of the natural language and from the perspective of the SQL query generated therefrom. Resolving the inherent ambiguities in natural language input includes “context-specific ambiguity {resolving ambiguous terminology...}” of natural-language query {the context-specific query...} due to “a term having different meanings depending on the query context {to context-specific terminology...}, the data domain {...associated with the software context of the computer application}, and the user goals” where, as explained with relation to an example, the word “top” includes both user intent context and software domain context in determining that “the query “Return the top movie” on a movie database, “top” may mean based on the number of ratings collected” and based on “SQL...[having] a strict syntax” the system resolves “the query ‘Return the movie with the best rating’... to a nested SQL query,” where resolving ambiguous terminology to context-specific terminology for text-to-SQL includes resolve the “vocabulary gap” (e.g., that the word “actress” resolves to “Actor.name”, where actress is ambiguous in the context of a database which does not rely on gender specific descriptors) and “schema ambiguity” (e.g., logically resolving whether “model” refers to “car.model” or “engine.model”).; Katsogiannis, ¶ p. 907-908, sections 2.1 and 2.2).
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the two-stage retrieval augmented generation of Subbiah to incorporate the teachings of Katsogiannis to include mapping, using a context mapping model, the terminology to a software-specific domain by resolving ambiguous terminology in the contextual query to context-specific terminology associated with the software context of the computer application. Subbiah discloses modification of a semantic query prior to generating the embedding (steps 304 and 306) to improve semantic search. Katsogiannis discloses that, when processing natural language for database queries, resolving ambiguity involves modifying the query string itself (via input serialization) to append schema-linking tokens or database-specific values, aligning the terminology with the database schema prior to generating the embeddings. It would have been obvious to modify the disclosure of Subbiah with the schema aligned input serialization taught by Katsogiannis, as part of a pre-processing step of step 304, to tag and align ambiguous terminology with strict relational database schema of the BDP structured data, thereby enriching the semantic vectors generated in step 306 and ensuring the downstream SQL query engine formulates highly accurate, error-free database commands, as recognized by Katsogiannis. (Katsogiannis, p. 907, col. 1, lines 1-18).
Regarding claim 16, the rejection of claim 15 is incorporated. Subbiah and Katsogiannis disclose all of the elements of the current invention as stated above. Subbiah further discloses wherein the one or more processors are further configured to cause the system to extract the plurality of data segment embeddings by: dividing content of the stored digital documents into the semantically consistent segments (“an optional information retrieval strategy is utilized to optimize information retrieval (e.g., in order to conduct a reliable semantic search)” which “may include the use of: (a) a chunking method (e.g., a method to chunk a document into text segments)”; Subbiah, ¶ [0110]) by determining content portions associated with one or more semantic concepts (“Optimization strategies” can further include “(b) a similarity score (e.g., utilized for measuring the text relevancy for the information retrieval)” and “(d) a searching strategy (e.g., a method to determining where to search and what chunks to retrieve); Subbiah, ¶ [0110]); and utilize the context retrieval model to encode a data segment embedding from a semantically consistent segment (“Optimization strategies” can further include “(c) retrievers (e.g., an algorithm utilized for retrieving relevant texts searching strategy”; Subbiah, ¶ [0110]).
Claims 3 is/are rejected under 35 U.S.C. 103 as being unpatentable over Subbiah and Katsogiannis as applied to claim 1 above, and further in view of Kar.
Regarding claim 3, the rejection of claim 1 is incorporated. Subbiah and Katsogiannis disclose all of the elements of the current invention as stated above. Subbiah further discloses wherein the context retrieval model is a…[machine] learning model (Embedding model is “trained based on construction data and the response identifies the semantically relevant section(s)” which is a machine learning model.; Subbiah, ¶ [0029], [0105]-[0106], [0108], [0113], [0120]) trained to determine, from among help documents, community question records, and help video transcripts, relevant data segments corresponding to the software contexts of the plurality of computer applications (the system can “allow customers to ask questions and extract insights from their documents as well as project data such as schedule, issues, RFIs, submittals, etc.” and discloses iterative refinement using documents such as “project performance data, such as issues, RFIs, specifications, takeoff data, locations data, submittals, specifications, schedule, change orders, cost and feedback loops to enhance the models’ understanding and predictive capabilities” where specifications, takeoff data, and change orders are understood as help documents; and issues, RFIs (Requests for Information), and feedback loops are understood as community question records.; Subbiah, ¶ [0023], [0029]). However, Subbiah fails to expressly recite wherein the context retrieval model is a contrastive learning model.
Kar teaches systems and methods for “contrastive learning of contextual retrieval augmented generation using semi-supervised retriever.” (Kar, ¶ [0001]). Regarding claim 3, Kar teaches wherein the context retrieval model is a contrastive learning model (discloses the source data for training as “A large corpus of unlabeled text data containing questions and their corresponding contexts” which comprises user questions and associated context is stored user interaction logs and stored digital documents, and the system generates a “plurality of positive and negative query-context pairs for the unlabeled text and the corresponding context”; Kar, ¶ [0031], [0051]).
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the two-stage retrieval augmented generation of Subbiah, as modified by the ambiguity resolving pre-processing of Katsogiannis, to incorporate the teachings of Kar to include wherein the context retrieval model is a contrastive learning model. “Self-supervised learning, when combined with retrieval-based architectures like Dense Passage Retriever,” as described in Kar, “enables fast and efficient retrieval of relevant passages, improving system response times and user experience,” which allows the system to reduce hallucinations and improve “retrieval speed without sacrificing the quality of results” generated by the generative model, as recognized by Kar. (Kar, ¶ [0011], [0094]-[0095]).
Claims 4 and 20 is/are rejected under 35 U.S.C. 103 as being unpatentable over Subbiah and Katsogiannis as applied to claims 1 and 15 above, and further in view of Joachims.
Regarding claim 4, the rejection of claim 1 is incorporated. Subbiah and Katsogiannis disclose all of the elements of the current invention as stated above. Subbiah further discloses wherein the context retrieval model is trained by: identifying a set of digital documents associated with client device interaction in relation to a sample query (the system can “allow customers to ask questions and extract insights from their documents as well as project data such as schedule, issues, RFIs, submittals, etc.” and discloses iterative refinement using documents including “project performance data, such as issues, RFIs, specifications, takeoff data, locations data, submittals, specifications, schedule, change orders, cost and feedback loops to enhance the models’ understanding and predictive capabilities” each of which are digital documents which are associated with a client device interaction, and as they are related to iterative refinement using “the wealth of data generated during construction projects” the interaction is with relation to a sample query.; Subbiah, ¶ [0023], [0029]). However, Subbiah fails to expressly recite determining, for a sample digital document from among the set of digital documents, a relevance metric in relation to the sample query based on comparing a selection total of the sample digital document and selection totals of other sample documents in the set of digital documents; and modifying parameters of the context retrieval model using a cross entropy loss function weighted according to the relevance metric.
Joachims teaches systems and methods for “Propensity-Weighted Ranking SVM for discriminative learning from implicit feedback,. (Joachims, Abstract). Regarding claim 4, Joachims teaches determining, for a sample digital document from among the set of digital documents, a relevance metric in relation to the sample query (Discloses training on “Implicit feedback (e.g., clicks, dwell times, etc.)” in the context of a user submitting a query and interacting with a ranked list of documents, based on a “probability value” referred to “as the propensity of the observation” based on interactions with “search rankings” “search results” and a “given query”; Joachims, ¶ Abstract, p. 783, col. 1, para. 3) based on comparing a selection total of the sample digital document and selection totals of other sample documents in the set of digital documents (generating the propensity estimation includes swapping “the result at rank k with the result at rank r” and then calculating the ratio of “observed click-through rates (CTR)” where the ratio of CTR is the claimed comparison of selection totals.; Joachims, ¶ p. 785, col. 1, paras. 2-3); and modifying parameters of the context retrieval model using a cross entropy loss function weighted according to the relevance metric (Further discloses “Empirical Risk Minimization (ERM)” disclosing “modifying parameters” by minimizing a “loss function,” where the loss function is explicitly weighted by the propensity to “derive a Propensity-Weighted Ranking SVM for discriminative learning from implicit feedback” and authors explicitly consider a cross entropy loss function in their future directions (a “propensity ERM approach” can be developed for “pointwise LTR” and “listwise LTR,” where listwise LTR incorporate cross entropy loss as the loss function.; Joachims, ¶ abstract; p. 782, col. 1, para. 4, p. 788, col. 2, para. 6).
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the two-stage retrieval augmented generation of Subbiah, as modified by the ambiguity resolving pre-processing of Katsogiannis, to incorporate the teachings of Joachims to include determining, for a sample digital document from among the set of digital documents, a relevance metric in relation to the sample query based on comparing a selection total of the sample digital document and selection totals of other sample documents in the set of digital documents; and modifying parameters of the context retrieval model using a cross entropy loss function weighted according to the relevance metric. Joachims discloses a “Propensity-Weighted Ranking SVM for discriminative learning from implicit feedback” which “allows training of ranking functions even in settings where queries do not repeat” and the authors show “empirically that the proposed learning method is highly effective in dealing with biases, that it is robust to noise and propensity model misspecification, and that it scales efficiently” while real-world analysis shows substantial improvement in retrieval performance, as recognized by Joachims. (Joachims, Abstract).
Regarding claim 20, the rejection of claim 15 is incorporated. Subbiah and Katsogiannis disclose all of the elements of the current invention as stated above. Subbiah further discloses wherein the context retrieval model is trained by: identifying a set of digital documents selected via client device interaction in relation to a sample query (the system can “allow customers to ask questions and extract insights from their documents as well as project data such as schedule, issues, RFIs, submittals, etc.” and discloses iterative refinement using documents including “project performance data, such as issues, RFIs, specifications, takeoff data, locations data, submittals, specifications, schedule, change orders, cost and feedback loops to enhance the models’ understanding and predictive capabilities” each of which are digital documents which are selected by the system for the iterative refinement via a client device interaction, and as they are related to iterative refinement using “the wealth of data generated during construction projects,” the interaction is with relation to a sample query.; Subbiah, ¶ [0023], [0029]). However, Subbiah fail(s) to expressly recite determining, for a sample digital document from among the set of digital documents, a relevance metric in relation to the sample query based on comparing a selection total of the sample digital document and selection totals of other sample documents in the set of digital documents; and modifying parameters of the context retrieval model using a cross entropy loss function weighted according to the relevance metric.
The relevance of Joachims is described above with relation to claim 4. Regarding claim 20, Joachims teaches determining, for a sample digital document from among the set of digital documents, a relevance metric in relation to the sample query (Discloses training on “Implicit feedback (e.g., clicks, dwell times, etc.)” in the context of a user submitting a query and interacting with a ranked list of documents, based on a “probability value” referred to “as the propensity of the observation” based on interactions with “search rankings” “search results” and a “given query”; Joachims, ¶ Abstract, p. 783, col. 1, para. 3) based on comparing a selection total of the sample digital document and selection totals of other sample documents in the set of digital documents (generating the propensity estimation includes swapping “the result at rank k with the result at rank r” and then calculating the ratio of “observed click-through rates (CTR)” where the ratio of CTR is the claimed comparison of selection totals.; Joachims, ¶ p. 785, col. 1, paras. 2-3); and modifying parameters of the context retrieval model using a cross entropy loss function weighted according to the relevance metric (Further discloses “Empirical Risk Minimization (ERM)” disclosing “modifying parameters” by minimizing a “loss function,” where the loss function is explicitly weighted by the propensity to “derive a Propensity-Weighted Ranking SVM for discriminative learning from implicit feedback” and authors explicitly consider a cross entropy loss function in their future directions (a “propensity ERM approach” can be developed for “pointwise LTR” and “listwise LTR,” where listwise LTR incorporate cross entropy loss as the loss function.; Joachims, ¶ abstract; p. 782, col. 1, para. 4, p. 788, col. 2, para. 6).
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the two-stage retrieval augmented generation of Subbiah, as modified by the ambiguity resolving pre-processing of Katsogiannis, to incorporate the teachings of Joachims to include determining, for a sample digital document from among the set of digital documents, a relevance metric in relation to the sample query based on comparing a selection total of the sample digital document and selection totals of other sample documents in the set of digital documents; and modifying parameters of the context retrieval model using a cross entropy loss function weighted according to the relevance metric. Joachims discloses a “Propensity-Weighted Ranking SVM for discriminative learning from implicit feedback” which “allows training of ranking functions even in settings where queries do not repeat” and the authors show “empirically that the proposed learning method is highly effective in dealing with biases, that it is robust to noise and propensity model misspecification, and that it scales efficiently” while real-world analysis shows substantial improvement in retrieval performance, as recognized by Joachims. (Joachims, Abstract).
Claims 5-7, and 17-18 is/are rejected under 35 U.S.C. 103 as being unpatentable over Subbiah and Katsogiannis as applied to claims 1 and 15 above, and further in view of Chen.
Regarding claim 5, the rejection of claim 1 is incorporated. Subbiah and Katsogiannis disclose all of the elements of the current invention as stated above. Subbiah further discloses further comprising: determining a relevant question-answer pair in relation to the contextual query from among a plurality of question-answer pairs generated utilizing a large language model (Discloses that the user query into “search vectors for a semantic search,” where search vectors are searched against stored vectors, and “an exemplary vector store may include the issues vector database 608A, a specs vector database 608B, and a submittals vector database 608C (collectively referred to as vector store/database 608),” and where “the semantic search is performed within each of the multiple different sections based on the search vectors and the stored vectors. The semantic search identifies the semantically relevant sections of the multiple different sections” and “vector store may be utilized as a dedicated repository for storing embedding vectors that are associated with specification chunks”; Subbiah, ¶ [0111], [0122]-[0123]) tuned based on the stored digital documents, wherein the stored digital documents comprise help documents, community question records, and... [unstructured data] (the system can “allow customers to ask questions and extract insights from their documents as well as project data such as schedule, issues, RFIs, submittals, etc.” and discloses iterative refinement using documents such as “project performance data, such as issues, RFIs, specifications, takeoff data, locations data, submittals, specifications, schedule, change orders, cost and feedback loops to enhance the models’ understanding and predictive capabilities” where specifications, takeoff data, and change orders are understood as help documents; and issues, RFIs (Requests for Information), and feedback loops are understood as community question records. Further discloses a process including “asking and answering of complex questions directly within construction documents” and “the process is bolstered by custom preprocessing algorithms designed to extract and index information” such as from “unstructured document”; Subbiah, ¶ [0023], [0029]); and generating the contextual response utilizing the response generator model prompted using the relevant question-answer pair (“an LLM is utilized to generate a response based on the contextual data prompt,” where the contextual data prompt includes “the semantically relevant sections of the multiple different sections (i.e., the semantic search filters the data in the data source to improve the search quality and accuracy)” and the embeddings of “the pre-processed text from the query”; Subbiah, ¶ [0108], [0111], [0113]). However, Subbiah fails to expressly recite wherein the unstructured data includes help video transcripts.
Chen teaches systems and methods for the fine tuning of machine learning models. (Chen, ¶ [0001]). Regarding claim 5, Chen teaches wherein the unstructured data includes help video transcripts (Discloses that the “domain specific training document is an unstructured data” which can include “video files”, where video files includes help video files, and where “machine-generated text of a content type” is “extracted from the domain-specific training document.” Thus unstructured data also includes help video transcripts.; Chen, ¶ [0132], [0139]).
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the two-stage retrieval augmented generation of Subbiah, as modified by the ambiguity resolving pre-processing of Katsogiannis, to incorporate the teachings of Chen to include wherein the unstructured data includes help video transcripts. Chen discloses “Fine-tuning a pre-trained machine learning model” and, more specifically, “adjusting the hyperparameters of the machine learning model that has been pre-trained on domain-neutral data” to adapt the “pre-trained machine learning model to perform a similar task in a domain-specific environment,” which improves performance regarding the domain-specific task and reduces hallucinations, while also “conserv[ing] computing resources such as power and memory,” as recognized by Chen. (Chen, ¶ [0020]-[0023]).
Regarding claim 6, the rejection of claim 1 is incorporated. Subbiah and Katsogiannis disclose all of the elements of the current invention as stated above. Subbiah further discloses wherein the response generator model comprises parameters learned by: generating a question-answer pair... utilizing a large language model (“an LLM is utilized to generate a response based on the contextual data prompt” where the contextual data prompt comprises “the semantically relevant sections (identified in the semantic search) are consolidated/aggregated (e.g., based on relevancy) into a contextual data prompt” alongside the user query, where the generative model is an LLM “trained {to learn...}based on construction data and the response identifies the semantically relevant section(s) {relevant digital documents for the respective software contexts of the computer applications}”; Subbiah, ¶ [0112]-[0113]) tuned based on the stored digital documents, wherein the stored digital documents comprise help documents, community question records, and… [unstructured data] (the system can “allow customers to ask questions and extract insights from their documents as well as project data such as schedule, issues, RFIs, submittals, etc.” and discloses iterative refinement using documents such as “project performance data, such as issues, RFIs, specifications, takeoff data, locations data, submittals, specifications, schedule, change orders, cost and feedback loops to enhance the models' understanding and predictive capabilities” where specifications, takeoff data, and change orders are understood as help documents; and issues, RFIs (Requests for Information), and feedback loops are understood as community question records. Further discloses a process including “asking and answering of complex questions directly within construction documents” and “the process is bolstered by custom preprocessing algorithms designed to extract and index information” such as from “unstructured document”; Subbiah, ¶ [0023], [0027], [0029]). However, Subbiah fails to expressly recite generating a question-answer pair from a ground truth document utilizing a large language model tuned based on the stored digital documents; generating a predicted response for a sample question from the question-answer pair utilizing the response generator model prompted by the sample question, the ground truth document, and a negative document; and refining the parameters based on a comparison of the predicted response to an answer corresponding to the sample question within the question-answer pair; and wherein the unstructured data includes help video transcripts.
The relevance of Chen is described above with relation to claim 5. Regarding claim 6, Chen teaches generating a question-answer pair from a ground truth document utilizing a large language model tuned based on the stored digital documents (Discloses a training data generator to generate input-output pairs, where the “training component 152 uses the manual input-output pairs 104 to train 154 the training data generator 156” as well as obtained “content items 108 by querying the storage system 140, and receiving, from the storage system 140, content items 160” and, in some examples, “the training data generator 156 trained by the training component 152 is an LLM”; Chen, ¶ [0040]-[0041]); generating a predicted response for a sample question from the question-answer pair utilizing the response generator model prompted by the sample question, the ground truth document, and a negative document (“In the second stage of the training pipeline, a second machine learning model” also referred to as “pretrained machine learning model 408,” is “fine-tuned to perform one or more domain-specific tasks using the generated supplemental training data” where the generated supplemental training data includes “An input-output pair (e.g., training input 402 and corresponding pseudo label 418, determined by the training data generator 256 described in FIG. 2)” which is “an input with an associated output” and includes “a predicted output and an actual output”; Chen, ¶ [0025], [0060], [0062], [0064]); and refining the parameters based on a comparison of the predicted response to an answer corresponding to the sample question within the question-answer pair (“During each training iteration, the weights are tuned to reduce the amount of error thereby minimizing the differences between (or otherwise converging) a predicted output and an actual output” and “the fine-tuning manager 430 fine-tunes the weights in the pretrained machine learning model 408” where “the value of the pretrained weights in the pretrained weight matrix is adjusted according to an error (e.g., the error 412 determined by the comparator 410 comparing the pseudo label 418 to the predicted output 406).”; Chen, ¶ [0060], [0070]); and wherein the unstructured data includes help video transcripts (Discloses that the “domain specific training document is an unstructured data” which can include “video files”, where video files includes help video files, and where “machine-generated text of a content type” is “extracted from the domain-specific training document.” Thus unstructured data also includes help video transcripts.; Chen, ¶ [0132], [0139]).
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the two-stage retrieval augmented generation of Subbiah, as modified by the ambiguity resolving pre-processing of Katsogiannis, to incorporate the teachings of Chen to include generating a question-answer pair from a ground truth document utilizing a large language model tuned based on the stored digital documents; generating a predicted response for a sample question from the question-answer pair utilizing the response generator model prompted by the sample question, the ground truth document, and a negative document; and refining the parameters based on a comparison of the predicted response to an answer corresponding to the sample question within the question-answer pair; and wherein the unstructured data includes help video transcripts. Chen discloses “Fine-tuning a pre-trained machine learning model” and, more specifically, “adjusting the hyperparameters of the machine learning model that has been pre-trained on domain-neutral data” to adapt the “pre-trained machine learning model to perform a similar task in a domain-specific environment,” which improves performance regarding the domain-specific task and reduces hallucinations, while also “conserv[ing] computing resources such as power and memory,” as recognized by Chen. (Chen, ¶ [0020]-[0023]).
Regarding claim 7, the rejection of claim 1 is incorporated. Subbiah and Katsogiannis disclose all of the elements of the current invention as stated above. However, Subbiah fails to expressly recite wherein the response generator model comprises parameters learned by: generating a negative response for a sample question utilizing the response generator model prompted by the sample question and a negative document, wherein the negative response indicates that no answer is found within the stored digital documents; and refining the parameters based on the negative response.
The relevance of Chen is described above with relation to claim 5. Regarding claim 7, Chen teaches wherein the response generator model comprises parameters learned by: generating a negative response for a sample question utilizing the response generator model prompted by the sample question and a negative document, (“the fine-tuning manager 430 can fine-tune the pretrained machine learning model 408 using negative samples. In some embodiments, negative training samples are obtained by the training data generator 156. For example, negative samples may be part of the input-output training pairs that are provided to fine-tune the pretrained machine learning model 408.”; Chen, ¶ [0083]) wherein the negative response indicates that no answer is found within the stored digital documents (“Fine-tuning the pretrained machine learning model 408 using negative samples allows the fine-tuned machine learning model 425 to identify when there is not a concept to be extracted from a sample. For example, given the phrase “today is a sunny day,” and the task to extract a skill, the fine-tuned machine learning model 425 would return “None” because there is no skill to be extracted from the phrase ‘today is a sunny day’.”; Chen, ¶ [0083]); and refining the parameters based on the negative response (The pretrained machine learning model is finetuned {refining the parameters} based on the negative samples {the negative response}; Chen, ¶ [0083]).
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the two-stage retrieval augmented generation of Subbiah, as modified by the ambiguity resolving pre-processing of Katsogiannis, to incorporate the teachings of Chen to include wherein the response generator model comprises parameters learned by: generating a negative response for a sample question utilizing the response generator model prompted by the sample question and a negative document, wherein the negative response indicates that no answer is found within the stored digital documents; and refining the parameters based on the negative response. Chen discloses “Fine-tuning a pre-trained machine learning model” and, more specifically, “adjusting the hyperparameters of the machine learning model that has been pre-trained on domain-neutral data” to adapt the “pre-trained machine learning model to perform a similar task in a domain-specific environment,” which improves performance regarding the domain-specific task and reduces hallucinations, while also “conserv[ing] computing resources such as power and memory,” as recognized by Chen. (Chen, ¶ [0020]-[0023]).
Regarding claim 17, the rejection of claim 15 is incorporated. Subbiah and Katsogiannis disclose all of the elements of the current invention as stated above. However, Subbiah fails to expressly recite wherein the one or more processors are further configured to cause the system to generate the semantically consistent segments by removing personally identifiable information from the stored digital documents utilizing a named entity removal model to replace the personally identifiable information with sanitized text.
The relevance of Chen is described above with relation to claim 5. Regarding claim 17, Chen teaches wherein the one or more processors are further configured to cause the system to generate the semantically consistent segments by removing personally identifiable information from the stored digital documents (“According to the techniques described herein, personal data may be removed from any training dataset that is used to train AI models. The techniques described herein may utilize tools for anonymizing member and customer data. For example, user’s personal data may be redacted and minimized in training datasets for training AI models through delexicalization tools and other privacy enhancing tools for safeguarding user data. The techniques described herein may minimize use of any personal data in training AI models, including removing and replacing personal data.”; Chen, ¶ [0162]) utilizing a named entity removal model to replace the personally identifiable information with sanitized text (Discloses delexicalization including the removal and replacing of “user’s personal data” with sanitized text, where named entity removal is a type of delexicalization, and where said data includes “personal data associated with a user, such as personal information provided by the user to the platform,” which is understood to include a user’s name, a user’s address and other information.; Chen, ¶ [0161], [0162]).
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the two-stage retrieval augmented generation of Subbiah, as modified by the ambiguity resolving pre-processing of Katsogiannis, to incorporate the teachings of Chen to include wherein the one or more processors are further configured to cause the system to generate the semantically consistent segments by removing personally identifiable information from the stored digital documents utilizing a named entity removal model to replace the personally identifiable information with sanitized text. Chen discloses “Fine-tuning a pre-trained machine learning model” and, more specifically, “adjusting the hyperparameters of the machine learning model that has been pre-trained on domain-neutral data” to adapt the “pre-trained machine learning model to perform a similar task in a domain-specific environment,” which improves performance regarding the domain-specific task and reduces hallucinations, while also “conserv[ing] computing resources such as power and memory,” as recognized by Chen. (Chen, ¶ [0020]-[0023]).
Regarding claim 18, the rejection of claim 15 is incorporated. Subbiah and Katsogiannis disclose all of the elements of the current invention as stated above. Subbiah further discloses wherein the one or more processors are further configured to cause the system to extract the plurality of data segment embeddings by extracting semantically consistent segments from: primary sources comprising document summaries of help documents (Discloses the presentation of “summarized responses with identified sources” where said responses are saved with the vector representation as part of the stored data which is searched based on the stored vectors, and “answers may be generated and tasks summarized based on the data queried and saved” where, in one example, “RFIs” are summarized as “RFI 2—'Missing soffit detail’—Status: Open; RFI 20—'Finish floor elevation’—Status: Open; RFI 60—Basement Slab Level—Status: Open” where RFIs are help document and the RFI title with the description of the RFI is a document summary; Subbiah, ¶ [0056], [0061], [0072]-[0075], [0107], [0109]); and derived sources comprising question-answer pairs generated from... [unstructured data] and community question records (the system can “allow customers to ask questions and extract insights from their documents as well as project data such as schedule, issues, RFIs, submittals, etc.” and discloses iterative refinement using documents such as “project performance data, such as issues, RFIs, specifications, takeoff data, locations data, submittals, specifications, schedule, change orders, cost and feedback loops to enhance the models’ understanding and predictive capabilities” where specifications, takeoff data, and change orders are understood as help documents; and issues, RFIs (Requests for Information), and feedback loops are understood as community question records. Further discloses a process including “asking and answering of complex questions directly within construction documents” and “the process is bolstered by custom preprocessing algorithms designed to extract and index information” such as from “unstructured document”; Subbiah, ¶ [0023], [0027], [0029]). However, Subbiah fails to expressly recite wherein the unstructured data includes help video transcripts.
The relevance of Chen is described above with relation to claim 5. Regarding claim 18, Chen teaches wherein the unstructured data includes help video transcripts (Discloses that the “domain specific training document is an unstructured data” which can include “video files”, where video files includes help video files, and where “machine-generated text of a content type” is “extracted from the domain-specific training document.” Thus unstructured data also includes help video transcripts.; Chen, ¶ [0132], [0139]).
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the two-stage retrieval augmented generation of Subbiah, as modified by the ambiguity resolving pre-processing of Katsogiannis, to incorporate the teachings of Chen to include wherein the unstructured data includes help video transcripts. Chen discloses “Fine-tuning a pre-trained machine learning model” and, more specifically, “adjusting the hyperparameters of the machine learning model that has been pre-trained on domain-neutral data” to adapt the “pre-trained machine learning model to perform a similar task in a domain-specific environment,” which improves performance regarding the domain-specific task and reduces hallucinations, while also “conserv[ing] computing resources such as power and memory,” as recognized by Chen. (Chen, ¶ [0020]-[0023]).
Claims 9, 11, and 14 is/are rejected under 35 U.S.C. 103 as being unpatentable over Subbiah in view of Chen and Siriwardhana.
Regarding claim 9, Subbiah discloses A non-transitory computer readable medium storing executable instructions which, when executed by a processing device, cause the processing device to perform operations (Systems and methods described with reference to “using generative AI to process user queries in the construction domain” which may be “tangibly embodied in a non-transitory computer-readable medium” as implemented using a “computer 902” including a “processor 904” and a “memory 906” where “the operating system 908 and the computer program 910 are comprised of computer program 910 instructions which, when accessed, read and executed by the computer 902, cause the computer 902 to perform the steps necessary to implement and/or use the present invention”; Subbiah, ¶ [0105], [0140], [0147]) comprising: [using]... a training dataset... relating to respective software contexts of computer applications; modifying... parameters of a context retrieval model to learn relevant digital documents for the respective software contexts of the computer applications (Embedding model as directed to “process[ing] user queries in the construction domain...within a construction software system” are fine-tuned and trained “to become more attuned to the construction industry’s evolving landscape” where the embedding model is “trained {to learn...}based on construction data and the response identifies the semantically relevant section(s) {relevant digital documents for the respective software contexts of the computer applications}”; Subbiah, ¶ [0029], [0105]-[0106], [0108], [0113], [0120]); generating, utilizing a response generator…, a predicted response… (“an LLM is utilized to generate a response based on the contextual data prompt” where the contextual data prompt comprises “the semantically relevant sections (identified in the semantic search) are consolidated/aggregated (e.g., based on relevancy) into a contextual data prompt” alongside the user query; Subbiah, ¶ [0112]-[0113]); and modifying parameters of the response generator model (Generative model as directed to “process[ing] user queries in the construction domain...within a construction software system” are fine-tuned and trained “to become more attuned to the construction industry’s evolving landscape” where the generative mode, which may also be a LLM, “the LLM is trained based on construction data and the response identifies the semantically relevant section(s)”; Subbiah, ¶ [0029], [0105]-[0106], [0108], [0113], [0120]). However, Subbiah fails to expressly recite comprising: generating a training dataset from stored user interaction logs and stored digital documents comprising question-answer pairs relating to respective software contexts of computer applications; modifying, using contrastive learning on the training dataset, parameters of a context retrieval model to learn relevant digital documents comprising relevant question-answer pairs for the respective software contexts of the computer applications; generating, utilizing a response generator model to process a sample query, a predicted response corresponding to the sample query from a ground truth digital document within the relevant digital documents, wherein the ground truth digital document comprises one or more question-answer pairs; and modifying parameters of the response generator model based on the predicted response to learn context-specific responses for the respective software contexts of the computer applications.
The relevance of Chen is described above with relation to claim 5. Regarding claim 9, Chen teaches generating, utilizing a response generator model to process a sample query, a predicted response corresponding to the sample query from a ground truth digital document within the relevant digital documents, (“In the second stage of the training pipeline, a second machine learning model” also referred to as “pretrained machine learning model 408,” is “fine-tuned to perform one or more domain-specific tasks using the generated supplemental training data” where the generated supplemental training data includes “An input-output pair (e.g., training input 402 and corresponding pseudo label 418, determined by the training data generator 256 described in FIG. 2)” which is “an input with an associated output” and includes “a predicted output and an actual output”; Chen, ¶ [0025], [0060], [0062], [0064]); and modifying parameters of the response generator model (“During each training iteration, the weights are tuned to reduce the amount of error thereby minimizing the differences between (or otherwise converging) a predicted output and an actual output.”; Chen, ¶ [0060]) based on the predicted response to learn context-specific responses for the respective software contexts of the computer applications (“the fine-tuning manager 430 fine-tunes the weights in the pretrained machine learning model 408” where “the value of the pretrained weights in the pretrained weight matrix is adjusted according to an error (e.g., the error 412 determined by the comparator 410 comparing the pseudo label 418 to the predicted output 406),” thus learning context-specific responses according to the adjustments for error. As understood in the context of Subbiah, said context specific responses are “for the respective construction software contexts of the computer applications.; Chen, ¶ [0070]).
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the two-stage retrieval augmented generation of Subbiah, to incorporate the teachings of Chen to include generating, utilizing a response generator model to process a sample query, a predicted response corresponding to the sample query from a ground truth digital document within the relevant digital documents; and modifying parameters of the response generator model based on the predicted response to learn context-specific responses for the respective software contexts of the computer applications. Chen discloses “Fine-tuning a pre-trained machine learning model” and, more specifically, “adjusting the hyperparameters of the machine learning model that has been pre-trained on domain-neutral data” to adapt the “pre-trained machine learning model to perform a similar task in a domain-specific environment,” which improves performance regarding the domain-specific task and reduces hallucinations, while also “conserv[ing] computing resources such as power and memory,” as recognized by Chen. (Chen, ¶ [0020]-[0023]). However, Subbiah and Chen fail to expressly recite generating a training dataset from stored user interaction logs and stored digital documents comprising question-answer pairs relating to respective software contexts of computer applications; modifying, using contrastive learning on the training dataset, parameters of a context retrieval model to learn relevant digital documents comprising relevant question-answer pairs for the respective software contexts of the computer applications wherein the ground truth digital document comprises one or more question-answer pairs.
Siriwardhana teaches RAG-end2end domain adaptation for domain-specific ODQA. (Siriwardhana, Abstract). Regarding claim 9, Siriwardhana teaches generating a training dataset from stored user interaction logs and stored digital documents comprising question-answer pairs relating to respective software contexts of computer applications (The system uses “the QAConv dataset (Wu et al., 2021b), which contains 35,000 QA pairs generated from 10,000 conversations that involved two or more parties. We use the train (25,000), valid (5,000), and test (5,000) splits given in the dataset to train and evaluate our model,” which relate to the specified domain (i.e., “finetuned only with domain-specific question-answer pairs”) where conversation logs are user interaction logs within the broadest reasonable interpretation. As well, conversation logs are the stored digital documents which “contains 35,000 QA pairs”; Siriwardhana, ¶ pg. 7, col. 1, lines 39-45, and col. 2, lines 40-44); modifying, using contrastive learning on the training dataset, parameters of a context retrieval model to learn relevant digital documents (“Our RAG-end2end training architecture” explicitly unfreezes and updates parameters of the retriever model, which is a “DPR… model pre-trained on Wikipedia-based question answering datasets” and comprising “two tower BERT-based networks: the Question Encoder (EQ) and the Passage Encoder (EP)” where the DPR is defined by training using contrastive learning. Specifically, the “similarity between a question (q) and a passage (p) is calculated by taking the dot product of the two embeddings” and the DPR calculates the similarity by contrasting positive query document pairings (maximizing the dot product between a query and the corresponding relevant document) against negative parings from the dataset (minimizing the dot product between the query and irrelevant documents).; Siriwardhana, ¶ pg. 4, FIG. 1, and col. 1, lines 6-16; pg. 6, col. 1, lines 28-44) comprising relevant question-answer pairs for the respective software contexts of the computer applications (The conversation QA comprises “only domain-specific question-answer pairs” which “contains 35,000 QA pairs” for the respective domain context, and where, as read in light of Subbiah, the domain refers to construction software context. It is further noted that the limitation to a software context is a field of use limitation which, without more, is not entitled to patentable weight. As explained, for example, in Siriwardhana, “RAG-end2end achieves significant performance improvements in all three domains” and alongside “several other experiments” which “validate our approach comprehensively... our results show that our approach is stable and generalizable across different domains.” One skilled in the art would readily understand the applicability to any domain, such as basketball, software, oil and gas exploration, and a host of others.; Siriwardhana, ¶ pg. 7, col. 1, lines 39-45, and col. 2, lines 40-44) wherein the ground truth digital document comprises one or more question-answer pairs (The disclosed retrieval model is retrieving within domain, based on the above training for domain adaptation and, as previously stated, contains “domain-specific question-answer pairs”; Siriwardhana, ¶ pg. 7, col. 1, lines 39-45, and col. 2, lines 40-44).
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the two-stage retrieval augmented generation of Subbiah, as modified by the machine learning model fine tuning of Chen, to incorporate the teachings of Siriwardhana to include generating a training dataset from stored user interaction logs and stored digital documents comprising question-answer pairs relating to respective software contexts of computer applications; modifying, using contrastive learning on the training dataset, parameters of a context retrieval model to learn relevant digital documents comprising relevant question-answer pairs for the respective software contexts of the computer applications wherein the ground truth digital document comprises one or more question-answer pairs. Subbiah discloses the overall system and method for generating answers to contextual queries within a closed domain, but lacks the explicit disclosure regarding the training of the various models. Siriwardhana teaches a specialized Retrieval augmented Generation (RAG) architecture to modify the parameters of the context retrieval model (e.g., the Dense Passage Retriever) using contrastive learning on a domain-specific training dataset to allow the retriever to learn relevant digital documents of a specialized domain. (Siriwardhana, Abstract). Siriwardhana further explains that standard retrieval models with frozen parameters struggle to retrieve highly relevant passages when applied to specialized closed domains. (Siriwardhana, pg. 1, col. 2, lines 1-10). It would have been obvious to one having ordinary skill in the art to modify Subbiah with the contrastive learning retrieval-update architecture of Siriwardhana prior to the effective filing data of the claimed invention, as this is the application of a known technique (Siriwardhana’s joint training of a retriever and generator on domain-specific dataset) to a known system (Subbiah’s document-based query answering architecture) to yield predictable results. The motivation for incorporating the teachings of Siriwardhana includes that such an incorporation would achieve the stated goal of “domain adaptation”, thereby providing significant performance improvements, higher retrieval accuracy, and a reduction in generative hallucinations, as compared to prior art systems, as recognized by Siriwardhana. (Siriwardhana, pg. 1, col. 2, lines 11-30; pg. 10, col. 2, lines 1-9)
Regarding claim 11, the rejection of claim 9 is incorporated. Subbiah, Chen, and Siriwardhana disclose all of the elements of the current invention as stated above. Subbiah further discloses wherein generating the training dataset comprises: generating question-answer pairs utilizing a large language model to process the stored digital documents, (“an LLM is utilized to generate a response based on the contextual data prompt” where the contextual data prompt comprises “the semantically relevant sections (identified in the semantic search) are consolidated/aggregated (e.g., based on relevancy) into a contextual data prompt” alongside the user query, where the generative model is an LLM “trained {to learn...}based on construction data and the response identifies the semantically relevant section(s) {relevant digital documents for the respective software contexts of the computer applications}”; Subbiah, ¶ [0112]-[0113]) wherein the stored digital documents comprise help documents, community question records, and [unstructured data] (the system can “allow customers to ask questions and extract insights from their documents as well as project data such as schedule, issues, RFIs, submittals, etc.” and discloses iterative refinement using documents such as “project performance data, such as issues, RFIs, specifications, takeoff data, locations data, submittals, specifications, schedule, change orders, cost and feedback loops to enhance the models’ understanding and predictive capabilities” where specifications, takeoff data, and change orders are understood as help documents; and issues, RFIs (Requests for Information), and feedback loops are understood as community question records. Further discloses a process including “asking and answering of complex questions directly within construction documents” and “the process is bolstered by custom preprocessing algorithms designed to extract and index information” such as from “unstructured document”; Subbiah, ¶ [0023], [0027], [0029]). However, Subbiah fails to expressly recite wherein the unstructured data includes help video transcripts; removing personally identifiable information from the community question records within the stored digital documents; and generating, from the help documents, document summaries comprising titles and descriptions of the help documents.
The relevance of Chen is described above with relation to claim 5. Regarding claim 11, Chen teaches wherein the unstructured data includes help video transcripts (Discloses that the “domain specific training document is an unstructured data” which can include “video files”, where video files includes help video files, and where “machine-generated text of a content type” is “extracted from the domain-specific training document.” Thus unstructured data also includes help video transcripts.; Chen, ¶ [0132], [0139]); removing personally identifiable information from the community question records within the stored digital documents (“According to the techniques described herein, personal data may be removed from any training dataset that is used to train AI models. The techniques described herein may utilize tools for anonymizing member and customer data. For example, user’s personal data may be redacted and minimized in training datasets for training AI models through delexicalization tools and other privacy enhancing tools for safeguarding user data. The techniques described herein may minimize use of any personal data in training AI models, including removing and replacing personal data.”; Chen, ¶ [0162]); and generating, from the help documents, document summaries comprising titles and descriptions of the help documents (“answers may be generated and tasks summarized based on the data queried and saved” where, in one example, “RFIs” are summarized as “RFI 2—'Missing soffit detail’—Status: Open; RFI 20—'Finish floor elevation’—Status: Open; RFI 60—Basement Slab Level—Status: Open” where RFI corresponds to a title and the description of the RFI is a description; Subbiah, ¶ [0061], [0072]-[0075]).
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the two-stage retrieval augmented generation of Subbiah, as modified by the contrastive learning retrieval-update architecture of Siriwardhana, to incorporate the teachings of Chen to include wherein the unstructured data includes help video transcripts; removing personally identifiable information from the community question records within the stored digital documents; and generating, from the help documents, document summaries comprising titles and descriptions of the help documents. Chen discloses “Fine-tuning a pre-trained machine learning model” and, more specifically, “adjusting the hyperparameters of the machine learning model that has been pre-trained on domain-neutral data” to adapt the “pre-trained machine learning model to perform a similar task in a domain-specific environment,” which improves performance regarding the domain-specific task and reduces hallucinations, while also “conserv[ing] computing resources such as power and memory,” as recognized by Chen. (Chen, ¶ [0020]-[0023]).
Regarding claim 14, the rejection of claim 9 is incorporated. Subbiah, Chen, and Siriwardhana disclose all of the elements of the current invention as stated above. Subbiah further discloses wherein modifying the parameters of the response generator model comprises: generating a question-answer pair... utilizing a large language model (“an LLM is utilized to generate a response based on the contextual data prompt” where the contextual data prompt comprises “the semantically relevant sections (identified in the semantic search) are consolidated/aggregated (e.g., based on relevancy) into a contextual data prompt” alongside the user query, where the generative model is an LLM “trained {to learn...}based on construction data and the response identifies the semantically relevant section(s) {relevant digital documents for the respective software contexts of the computer applications}”; Subbiah, ¶ [0112]-[0113]) tuned based on the stored digital documents, wherein the stored digital documents comprise help documents, community question records, and… [unstructured data] (the system can “allow customers to ask questions and extract insights from their documents as well as project data such as schedule, issues, RFIs, submittals, etc.” and discloses iterative refinement using documents such as “project performance data, such as issues, RFIs, specifications, takeoff data, locations data, submittals, specifications, schedule, change orders, cost and feedback loops to enhance the models’ understanding and predictive capabilities” where specifications, takeoff data, and change orders are understood as help documents; and issues, RFIs (Requests for Information), and feedback loops are understood as community question records. Further discloses a process including “asking and answering of complex questions directly within construction documents” and “the process is bolstered by custom preprocessing algorithms designed to extract and index information” such as from “unstructured document”; Subbiah, ¶ [0023], [0027], [0029]). However, Subbiah fails to expressly recite generating a question-answer pair from a ground truth document utilizing a large language model tuned based on the stored digital documents wherein the unstructured data includes help video transcripts; generating a predicted response for a sample question from the question-answer pair utilizing the response generator model prompted by the sample question, the ground truth document, and a negative document; and updating the parameters of the response generator model based on a comparison of the predicted response to an answer corresponding to the sample question within the question-answer pair.
The relevance of Chen is described above with relation to claim 5. Regarding claim 14, Chen teaches generating a question-answer pair from a ground truth document utilizing a large language model tuned based on the stored digital documents (Discloses a training data generator to generate input-output pairs, where the “training component 152 uses the manual input-output pairs 104 to train 154 the training data generator 156” as well as obtained “content items 108 by querying the storage system 140, and receiving, from the storage system 140, content items 160” and, in some examples, “the training data generator 156 trained by the training component 152 is an LLM”; Chen, ¶ [0040]-[0041]) wherein the unstructured data includes help video transcripts (Discloses that the “domain specific training document is an unstructured data” which can include “video files”, where video files includes help video files, and where “machine-generated text of a content type” is “extracted from the domain-specific training document.” Thus unstructured data also includes help video transcripts.; Chen, ¶ [0132], [0139]); generating a predicted response for a sample question from the question-answer pair utilizing the response generator model prompted by the sample question, the ground truth document, and a negative document (“In the second stage of the training pipeline, a second machine learning model” also referred to as “pretrained machine learning model 408,” is “fine-tuned to perform one or more domain-specific tasks using the generated supplemental training data” where the generated supplemental training data includes “An input-output pair (e.g., training input 402 and corresponding pseudo label 418, determined by the training data generator 256 described in FIG. 2)” which is “an input with an associated output” and includes “a predicted output and an actual output”; Chen, ¶ [0025], [0060], [0062], [0064]); and updating the parameters of the response generator model based on a comparison of the predicted response to an answer corresponding to the sample question within the question-answer pair (“During each training iteration, the weights are tuned to reduce the amount of error thereby minimizing the differences between (or otherwise converging) a predicted output and an actual output” and “the fine-tuning manager 430 fine-tunes the weights in the pretrained machine learning model 408” where “the value of the pretrained weights in the pretrained weight matrix is adjusted according to an error (e.g., the error 412 determined by the comparator 410 comparing the pseudo label 418 to the predicted output 406).”; Chen, ¶ [0060], [0070]).
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the two-stage retrieval augmented generation of Subbiah, as modified by the contrastive learning retrieval-update architecture of Siriwardhana, to incorporate the teachings of Chen to include generating a question-answer pair from a ground truth document utilizing a large language model tuned based on the stored digital documents wherein the unstructured data includes help video transcripts; generating a predicted response for a sample question from the question-answer pair utilizing the response generator model prompted by the sample question, the ground truth document, and a negative document; and updating the parameters of the response generator model based on a comparison of the predicted response to an answer corresponding to the sample question within the question-answer pair. Chen discloses “Fine-tuning a pre-trained machine learning model” and, more specifically, “adjusting the hyperparameters of the machine learning model that has been pre-trained on domain-neutral data” to adapt the “pre-trained machine learning model to perform a similar task in a domain-specific environment,” which improves performance regarding the domain-specific task and reduces hallucinations, while also “conserv[ing] computing resources such as power and memory,” as recognized by Chen. (Chen, ¶ [0020]-[0023]).
Claims 10 is/are rejected under 35 U.S.C. 103 as being unpatentable over Subbiah, Chen, and Siriwardhana as applied to claim 9 above, and further in view of Joachims.
Regarding claim 10, the rejection of claim 9 is incorporated. Subbiah, Chen, and Siriwardhana disclose all of the elements of the current invention as stated above. Subbiah further discloses wherein the operations further comprise: identifying a set of digital documents associated with client device interaction in relation to the sample query (the system can “allow customers to ask questions and extract insights from their documents as well as project data such as schedule, issues, RFIs, submittals, etc.” and discloses iterative refinement using documents including “project performance data, such as issues, RFIs, specifications, takeoff data, locations data, submittals, specifications, schedule, change orders, cost and feedback loops to enhance the models’ understanding and predictive capabilities” each of which are digital documents which are associated with a client device interaction, and as they are related to iterative refinement using “the wealth of data generated during construction projects” the interaction is with relation to a sample query.; Subbiah, ¶ [0023], [0029]). However, Subbiah fails to expressly recite determining, for a sample digital document from among the set of digital documents, a relevance metric based on comparing selections of the sample digital document and selections of other sample documents in the set of digital documents; and modifying parameters of the context retrieval model using a cross entropy loss function weighted according to the relevance metric.
The relevance of Joachims is described above with relation to claim 4. Regarding claim 10, Joachims teaches determining, for a sample digital document from among the set of digital documents, a relevance metric (Discloses training on “Implicit feedback (e.g., clicks, dwell times, etc.)” in the context of a user submitting a query and interacting with a ranked list of documents, based on a “probability value” referred to “as the propensity of the observation” based on interactions with “search rankings” “search results” and a “given query”; Joachims, ¶ Abstract, p. 783, col. 1, para. 3) based on comparing selections of the sample digital document and selections of other sample documents in the set of digital documents (generating the propensity estimation includes swapping “the result at rank k with the result at rank r” and then calculating the ratio of “observed click-through rates (CTR)” where the ratio of CTR is the claimed comparison of selection totals.; Joachims, ¶ p. 785, col. 1, paras. 2-3); and modifying parameters of the context retrieval model using a cross entropy loss function weighted according to the relevance metric (Further discloses “Empirical Risk Minimization (ERM)” disclosing “modifying parameters” by minimizing a “loss function,” where the loss function is explicitly weighted by the propensity to “derive a Propensity-Weighted Ranking SVM for discriminative learning from implicit feedback” and authors explicitly consider a cross entropy loss function in their future directions (a “propensity ERM approach” can be developed for “pointwise LTR” and “listwise LTR,” where listwise LTR incorporate cross entropy loss as the loss function.; Joachims, ¶ abstract; p. 782, col. 1, para. 4, p. 788, col. 2, para. 6).
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the two-stage retrieval augmented generation of Subbiah, as modified by the machine learning model fine tuning of Chen, as modified by the contrastive learning retrieval-update architecture of Siriwardhana, to incorporate the teachings of Joachims to include determining, for a sample digital document from among the set of digital documents, a relevance metric based on comparing selections of the sample digital document and selections of other sample documents in the set of digital documents; and modifying parameters of the context retrieval model using a cross entropy loss function weighted according to the relevance metric. Joachims discloses a “Propensity-Weighted Ranking SVM for discriminative learning from implicit feedback” which “allows training of ranking functions even in settings where queries do not repeat” and the authors show “empirically that the proposed learning method is highly effective in dealing with biases, that it is robust to noise and propensity model misspecification, and that it scales efficiently” while real-world analysis shows substantial improvement in retrieval performance, as recognized by Joachims. (Joachims, Abstract).
Claims 12 is/are rejected under 35 U.S.C. 103 as being unpatentable over Subbiah, Chen, and Siriwardhana as applied to claim 9 above, and further in view of Zhan.
Regarding claim 12, the rejection of claim 9 is incorporated. Subbiah, Chen, and Siriwardhana disclose all of the elements of the current invention as stated above. However, Subbiah fails to expressly recite wherein modifying the parameters of the context retrieval model comprises: determining a cosine similarity between a query embedding and a plurality of data segment embeddings extracted from the stored digital documents; and modifying the parameters using a mean squared error function based on the cosine similarity.
Zhan teaches systems and methods for “unsupervised cross-modal Hash retrieval training” (Zhan, ¶ [0001]). Regarding claim 12, Zhan teaches wherein modifying the parameters of the context retrieval model comprises: determining a cosine similarity between a query embedding and a plurality of data segment embeddings extracted from the stored digital documents (“Step 407, the first hash model and the second hash model respectively utilize the mode with small error selected by the other party to perform” backpropagation “and update the network parameters of the first hash model.”; Zhan, ¶ [0024]); and modifying the parameters using a mean squared error function based on the cosine similarity (“calculating cosine similarity on the feature representation, and calculating a mean square error with the pseudo label”; Zhan, ¶ [0024]).
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the two-stage retrieval augmented generation of Subbiah, as modified by the machine learning model fine tuning of Chen, as modified by the contrastive learning retrieval-update architecture of Siriwardhana, to incorporate the teachings of Zhan to include wherein modifying the parameters of the context retrieval model comprises: determining a cosine similarity between a query embedding and a plurality of data segment embeddings extracted from the stored digital documents; and modifying the parameters using a mean squared error function based on the cosine similarity. Zhan teaches “an unsupervised cross-modal hashing retrieval method based on noisy label learning. This invention can achieve this by setting up two dual hashing model groups, feeding each other relatively clean pseudo-labels, thus minimizing the model's learning from being misled by noisy pseudo-labels. Ultimately, the model converges to a better position, and its performance on the test dataset is better than other unsupervised cross-modal hashing methods,” as recognized by Zhan. (Zhan, ¶ [0017]).
Claims 13 is/are rejected under 35 U.S.C. 103 as being unpatentable over Subbiah, Chen, and Siriwardhana as applied to claim 9 above, and further in view of Xiong.
Regarding claim 13, the rejection of claim 9 is incorporated. Subbiah, Chen, and Siriwardhana disclose all of the elements of the current invention as stated above. However, Subbiah fails to expressly recite wherein the operations further comprise: determining, utilizing a contrastive learning model, a shared weight for a query embedding extracted from the sample query and a plurality of data segment embeddings extracted from data segments of the stored digital documents; and modifying parameters of the context retrieval model utilizing a loss function informed by the shared weight.
Xiong teaches systems and methods for machine learning using “Approximate nearest neighbor Negative Contrastive Learning (ANCE)”. (Xiong, Abstract). Regarding claim 13, Xiong teaches wherein the operations further comprise: determining, utilizing a contrastive learning model, a shared weight for a query embedding extracted from the sample query and a plurality of data segment embeddings extracted from data segments of the stored digital documents (Authors disclose “Approximate nearest neighbor Negative Contrastive Estimation, (ANCE)” which is contrastive learning, and “ANCE can be used to train any dense retrieval model” described here using a “BERT Siamese/Dual Encoder (shared between q and d),” where q is the query and d is the document, and comprising the shared weight; Xiong, ¶ pg. 4, lines. 22-33); and modifying parameters of the context retrieval model utilizing a loss function informed by the shared weight (“The learning of this representation often follows standard learning to rank (Liu, 2009): Given a query q, a set of relevant document D+ and irrelevant ones D−, find the best θ” can be determined based on equation 2, where, in some examples, “the loss l() can be binary cross entropy (BCE), hinge loss, or negative log likelihood (NLL).”; Xiong, ¶ pg. 2, lines 40-44; Eq. (2)).
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the two-stage retrieval augmented generation of Subbiah, as modified by the machine learning model fine tuning of Chen, as modified by the contrastive learning retrieval-update architecture of Siriwardhana, to incorporate the teachings of Xiong to include wherein the operations further comprise: determining, utilizing a contrastive learning model, a shared weight for a query embedding extracted from the sample query and a plurality of data segment embeddings extracted from data segments of the stored digital documents; and modifying parameters of the context retrieval model utilizing a loss function informed by the shared weight. Acknowledging that “dense retrieval” is a core component “to retrieval relevant information for …extractive/generative QA” and “fact verification” the Xiong indicates ANCE as an improvement on the training of retrieval models in such situations, explaining that “ANCE is orthogonal with those lines of research and focuses on the representation learning for dense retrieval” where the “effectiveness of ANCE on web search, question answering, and in a commercial search environment… nearly matches the accuracy of BERT-based cascade IR pipeline, while being 100x more efficient,” thus providing state of the art retrieval accuracy while reducing processing overhead and improving training convergence, among other benefits, as recognized by Xiong. (Xiong, pg. 8, para. 4).
Claims 19 is/are rejected under 35 U.S.C. 103 as being unpatentable over Subbiah and Katsogiannis as applied to claim 15 above, and further in view of Chen and Mahajan (U.S. Pat. App. Pub. No. 2021/0026820, hereinafter Mahajan).
Regarding claim 19, the rejection of claim 15 is incorporated. Subbiah and Katsogiannis disclose all of the elements of the current invention as stated above. Subbiah further discloses wherein the one or more processors are further configured to cause the system to: determine a set of question-answer pairs utilizing a large language model to process the stored digital documents, (Discloses that the user query into “search vectors for a semantic search,” where search vectors are searched against stored vectors, and “an exemplary vector store may include the issues vector database 608A, a specs vector database 608B, and a submittals vector database 608C (collectively referred to as vector store/database 608),” and where “the semantic search is performed within each of the multiple different sections based on the search vectors and the stored vectors. The semantic search identifies the semantically relevant sections of the multiple different sections” and “vector store may be utilized as a dedicated repository for storing embedding vectors that are associated with specification chunks”; Subbiah, ¶ [0111], [0122]-[0123]) wherein the stored digital documents comprise help documents, community question records, and [unstructured data] (the system can “allow customers to ask questions and extract insights from their documents as well as project data such as schedule, issues, RFIs, submittals, etc.” and discloses iterative refinement using documents such as “project performance data, such as issues, RFIs, specifications, takeoff data, locations data, submittals, specifications, schedule, change orders, cost and feedback loops to enhance the models’ understanding and predictive capabilities” where specifications, takeoff data, and change orders are understood as help documents; and issues, RFIs (Requests for Information), and feedback loops are understood as community question records. Further discloses a process including “asking and answering of complex questions directly within construction documents” and “the process is bolstered by custom preprocessing algorithms designed to extract and index information” such as from “unstructured document”; Subbiah, ¶ [0023], [0027], [0029]); and generate the contextual response utilizing the response generator model informed by the subset of question-answer pairs (“an LLM is utilized to generate a response based on the contextual data prompt,” where the contextual data prompt includes “the semantically relevant sections of the multiple different sections (i.e., the semantic search filters the data in the data source to improve the search quality and accuracy)” and the embeddings of “the pre-processed text from the query”; Subbiah, ¶ [0108], [0111], [0113]). However, Subbiah fails to expressly recite wherein the unstructured data includes help video transcripts; [and] generate, from the set of question-answer pairs, a subset of question-answer pairs by removing redundant question-answer pairs using a Levenshtein distance.
The relevance of Chen is described above with relation to claim 5. Regarding claim 19, Chen teaches wherein the unstructured data includes help video transcripts (Discloses that the “domain specific training document is an unstructured data” which can include “video files”, where video files includes help video files, and where “machine-generated text of a content type” is “extracted from the domain-specific training document.” Thus unstructured data also includes help video transcripts.; Chen, ¶ [0132], [0139]).
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the two-stage retrieval augmented generation of Subbiah, as modified by the ambiguity resolving pre-processing of Katsogiannis, to incorporate the teachings of Chen to include wherein the unstructured data includes help video transcripts. Chen discloses “Fine-tuning a pre-trained machine learning model” and, more specifically, “adjusting the hyperparameters of the machine learning model that has been pre-trained on domain-neutral data” to adapt the “pre-trained machine learning model to perform a similar task in a domain-specific environment,” which improves performance regarding the domain-specific task and reduces hallucinations, while also “conserv[ing] computing resources such as power and memory,” as recognized by Chen. (Chen, ¶ [0020]-[0023]). However, Subbiah and Chen fail to expressly recite generate, from the set of question-answer pairs, a subset of question-answer pairs by removing redundant question-answer pairs using a Levenshtein distance
Mahajan teaches systems and methods for removing redundant database entries. (Mahajan, ¶ [0001]). Regarding claim 19, Mahajan teaches generate, from the set of question-answer pairs, a subset of question-answer pairs by removing redundant question-answer pairs using a Levenshtein distance (“attribute similarities are computed… on each column and are computed across each pair of rows” including “fuzzy matching comparators…whereby the pair of rows match if they have low Levenshtein distance,” where matching entries are removed as duplicates, and though presented in the context of contact information, the reference clearly indicates that the disclosed embodiments are not limited to such.; Mahajan, ¶ [0018], [0033], [0055]).
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the two-stage retrieval augmented generation of Subbiah, as modified by the ambiguity resolving pre-processing of Katsogiannis, as modified by the machine learning model fine tuning of Chen, to incorporate the teachings of Mahajan to include generate, from the set of question-answer pairs, a subset of question-answer pairs by removing redundant question-answer pairs using a Levenshtein distance. As explained by Mahajan, deduplication of database entries is well known in the art for, at least, improving search efficiency and conserving storage space. In the context of Subbiah, removing duplicate entries through the use of well-known similarity measures, such as Levenshtein distance, would have been obvious to one of ordinary skill in the art and provide the above well-known benefits, which would speed the search and prompt generation processes, as well as helping to avoid overfitting the data and/or memorizing certain data points due to duplicate entries in the training set, as recognized in the context of Mahajan and Subbiah. (Mahajan, ¶ [0002]-[0004]).
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure.
Non-Patent Literature to Pan et al. (Pan, F., Canim, M., Glass, M., Gliozzo, A. and Hendler, J., 2022. End-to-end table question answering via retrieval-augmented generation. arXiv preprint arXiv:2203.16714.) discloses T-RAG, an end-to-end Table QA model, where a non-parametric dense vector index is fine-tuned jointly with BART, a parametric sequence-to-sequence model to generate answer tokens and incorporate a unified pipeline to automatically search through a table corpus to directly locate the correct answer from the table cells.
Applicant's amendment necessitated the new ground(s) of rejection presented in this Office action. Accordingly, THIS ACTION IS MADE FINAL. See MPEP § 706.07(a). Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a).
A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to Sean E. Serraguard whose telephone number is (313)446-6627. The examiner can normally be reached 07:00-17:00 M-F.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Daniel C. Washburn can be reached at (571) 272-5551. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of published or unpublished applications may be obtained from Patent Center. Unpublished application information in Patent Center is available to registered users. To file and manage patent submissions in Patent Center, visit: https://patentcenter.uspto.gov. Visit https://www.uspto.gov/patents/apply/patent-center for more information about Patent Center and https://www.uspto.gov/patents/docx for information about filing in DOCX format. For additional questions, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/Sean E Serraguard/Primary Examiner, Art Unit 2657
Prosecution Insights